GazePlotter v1.10.5, Nov 2026
Launch app App
Documentation Chapters

Scanpath Similarity

Scanpath similarity metrics score how alike every pair of participants looked. Levenshtein and Needleman-Wunsch compare the order of visited AOIs. Transition similarity compares which AOI-to-AOI moves they made. Heatmap overlap compares where they looked, in any order.

Output

All scanpath similarity metrics produce a participant-by-participant grid: one similarity score for every pair of participants in the selected group. This output is what the Scanpath Similarity visualization consumes; because the result is inherently a property of the whole group, it cannot be reshaped for other plots.

Invariants

  • Measurement class: Group-level. The result is inherently a property of a pair of participants, where both the row and column axes are the participants themselves. There is no per-participant value to combine across the group; the matrix is already the across-participant result, so no further aggregation applies.
  • Windowing: Not available: the comparison always covers the whole selected time range. Cropping by a Time of Interest (the Start/End bounds in the Pane) is supported; a fixation is encoded into the scanpath string only if it begins within the range. Transition similarity counts every fixation that overlaps the range, as the Transitions metrics do. Heatmap overlap counts the part of each fixation inside the range, like the Fixation Heatmap.

Scanpath String Encoding

To compare scanpaths, GazePlotter converts each participant’s fixation sequence into an AOI-letter string:

  • Each active AOI is mapped to a unique letter (the first AOI becomes ‘A’, the second ‘B’, and so on).
  • Fixations that do not land within any active AOI are encoded using a special hash character (#).

Metrics

Levenshtein similarity and Needleman-Wunsch similarity compare the scanpath strings of every pair of participants. Each resulting matrix is symmetric (similarity from participant i to j equals j to i) and has a diagonal of 1.0 (every participant is identical to themselves). A pair where both scanpaths are empty (no fixations in the analyzed range) carries no value, and so does the diagonal cell of a participant with an empty scanpath: the matrix renders those cells as missing, and such pairs never form edges or cliques in the ScanGraph view. A pair where only one scanpath is empty scores 0, as the formulas define.

Both have one parameter:

  1. Collapse consecutive AOIs:
    • Off (default): Consecutive fixations inside the same AOI are preserved. A participant who fixates AOI A three times before moving to B is encoded as “AAAB”.
    • On: Consecutive identical AOI entries are collapsed into a single letter (“AAAB” → “AB”). This removes duration-based variance, isolating the analysis to focus purely on the structural order of visited regions.

Levenshtein Similarity

Calculates the minimum edit distance (insertions, deletions, and substitutions) required to transform one scanpath string into the other. The score is normalized by the maximum length of the two scanpaths:

Similarity = 1 - (Edit Distance / max(Length_1, Length_2))

Needleman-Wunsch Similarity

A dynamic programming algorithm for global sequence alignment. It scores alignments with the ScanGraph weights (Match = +1, Mismatch = -1, Gap = 0) and divides the alignment score by the maximum length of the two scanpaths:

Similarity = Alignment Score / max(Length_1, Length_2)

This is the similarity parameter p of the ScanGraph method (Doležalová & Popelka, 2016; see Reference), so thresholds chosen here are directly comparable with results from the standalone ScanGraph tool.

Transition Similarity

Compares the mix of AOI transitions two participants made, using the counts of the Transitions metric. The score is the cosine of the two count matrices:

Similarity = Σ (Count_1 × Count_2) / √(Σ Count_1² × Σ Count_2²)

1 means the same mix of transitions, 0 means none in common. Only the mix counts, not how many transitions or when. Also in GazeAlytics (Chen et al., 2023).

  1. Count mode: Visit changes (default) counts moves between AOIs. Fixation pairs also counts staying in an AOI, which pulls the score towards dwell.

A participant without transitions has no value.

Heatmap Overlap

Compares where two participants looked, ignoring order. Needs fixation coordinates, not AOIs. Each participant gets a Fixation Heatmap of their share of fixation time, and the overlap sums the smaller share in every cell:

Similarity = Σ min(Share_1, Share_2)

1 means identical maps, 0 means no common area. This is SIM, or histogram intersection (Swain & Ballard, 1991; Bylinskii et al., 2019), as in GazeAlytics (Chen et al., 2023).

  1. Radius R (0 = Auto): Kernel radius, as on the Fixation Heatmap. A larger radius raises every value, so report it.

A participant with no located fixation in the time range has no value: their whole row and column stay empty, diagonal included.

Most viewers look at the same central areas, so values sit well above 0. Compare pairs with each other, not with 0; the ScanGraph’s Auto threshold does this for you. To see groups in the matrix, order it by similarity and raise the color scale’s Min.

Visualizing the Similarity Matrix

The resulting similarity scores can be explored in two ways under the Scanpath Similarity visualization:

  1. Heatmap Matrix: Displays the full participant-by-participant grid, colored by similarity strength, to identify clusters of participants with similar viewing strategies.
  2. ScanGraph Network: Represents participants as nodes in a graph. An edge is drawn between two participants only if their similarity score meets a threshold, helping visualize communities of shared attention.

Reference

The Needleman-Wunsch scoring, the similarity parameter p, and the ScanGraph network are based on:

Doležalová, J., & Popelka, S. (2016). ScanGraph: A novel scanpath comparison method using visualisation of graph cliques. Journal of Eye Movement Research, 9(4). https://doi.org/10.16910/jemr.9.4.5

Heatmap overlap and transition similarity follow:

Swain, M. J., & Ballard, D. H. (1991). Color indexing. International Journal of Computer Vision, 7(1), 11–32. https://doi.org/10.1007/BF00130487

Bylinskii, Z., Judd, T., Oliva, A., Torralba, A., & Durand, F. (2019). What do different evaluation metrics tell us about saliency models? IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(3), 740–757. https://doi.org/10.1109/TPAMI.2018.2815601

Chen, K.-T., Prouzeau, A., Langmead, J., Whitelock-Jones, R. T., Lawrence, L., Dwyer, T., Hurter, C., Weiskopf, D., & Goodwin, S. (2023). Gazealytics: A unified and flexible visual toolkit for exploratory and comparative gaze analysis. Proceedings of the 2023 Symposium on Eye Tracking Research and Applications (ETRA). https://doi.org/10.1145/3588015.3589844